Deep Anomaly Detection with Extended Transformer-Based Model on Tennessee Eastman Process Dataset

Fabian Schoch, Pascal Graf, Tobias Schmieg, Carsten Wittenberg, Carsten Lanquillon, Nicolaj C. Stache

Abstract: Identifying anomalies in multivariate time series data for modern industrial production processes is becoming increasingly essential to prevent system failures, reduce downtimes, and enhance safety. The advent of the Internet of Things has enabled the collection of vast amounts of data from industrial machinery, providing a rich source of information for anomaly detection. In this context, we propose a Transformer-based reconstruction approach with additional enhancements, including projection and patching, as well as a transposed Convolutional Neural Network for reconstruction. Our approach significantly outperforms existing non-Transformer-based models on the Tennessee Eastman Process dataset.

External IDs:dblp:conf/csit/SchochGSWLS24